Abstract: To improve the naturalness of head animation of virtual signers, this paper proposes neighborhood preserving canonical correlation analysis to realize head motion prediction. Firstly, synchronous hand and head motion features are extracted from a real signer performance. Secondly, the nonlinear canonical correlation analysis is applied to build a mapping from hand to head motion features. Meanwhile, neighborhood preserving constraint is employed to capture action transitions of successive motion frames, and better model the smoothness between them. Experimental results show that the proposed method can achieve more realistic and natural head animation of the signing avatar.